Validity of a machine learning estimation of blood volumes during altitude training

Abstract

Introduction

Altitude training is widely used in endurance sports to improve near sea-level performance through physiological adaptations. Overall gains of 1% total hemoglobin mass (Hbmass) for each 100 hours spent at altitude may be expected. However, since individual athlete response differs widely, an adequate assessment of actual Hbmass and plasma volume (PV) changes is paramount.

We recently proposed a machine learning model to estimate Hbmass and PV from a single blood sample (Moreillon et al., 2023). This study investigated the validity of the latter predictive model to detect actual Hbmass and PV adaptations during an altitude camp in professional cyclists.

Methods

Blood samples were taken at the start and the end of a 21-days altitude training camp (at an altitude of 2050 m (Kühtai, Austria) from 21 professional cyclists (25.4 ± 3.3 yrs, 69.8 ± 5.0 kg, maximal aerobic power 560 ± 34 W). A Co-rebreathing method was used to assess actual Hbmass and PV changes during the altitude sojourn. A machine learning prediction model trained on blood variables measured in >700 subjects was applied on individual complete blood count values to estimate Hbmass and PV as detailed elsewhere (Moreillon et al., 2023).

Results

At the end of the camp, a significant decrease in PV was measured (-167 ± 345 mL; p<0.05) whereas Hbmass significantly increased on average (+40 ± 36 g; p<0.001). Conversely, a significant increase in hematocrit (+2.7 + 2.2%pt ; p<0.001) and hemoglobin concentration ([Hb], +0.63 + 0.75 g×dL^-1; p<0.01). Predicted PV and Hbmass were both significantly correlated to measured PV and Hbmass (R²=0.53, p<0.001; R²=0.26; p<0.001; respectively) with an estimated PV decrease of -162 ± 232 mL. The root mean square error (RMSE) for the estimation between predicted and measured difference was of 95 mL. The model estimated an Hbmass increase of 7 ± 31 g with a RMSE of 20 g for individual differences between predicted and measured Hbmass.

Discussion/Conclusion

The observed results for hematocrit and [Hb] reflect the shifts in PV occuring during the hypoxic exposure. These variables are usually poor indicators of the effects of an altitude training as PV also rapidly shifts upon return to a lower altitude. Interestingly, the predictive model estimations of the downshift in PV observed at altitude were corresponding closely to the measured values.

Conversely, predicted values for Hbmass underestimated the actual gains, indicating that the predictive model may not be sensitive enough to discriminate actual variations due to a prolonged hypoxic expopsure. In conclusion, this study suggest that CO-rebreathing may definitely represent a gold-standard to asess the effectiveness of an altitude sojourn on increasing Hbmass while a predictive machine-learning model may allow to monitor PV shifts from a single blood sample to help decision-making regarding body fluid balance before, during and after an altitude training camp in professional cyclists.

References

Moreillon, B., Krumm, B., Saugy, J. J., Saugy, M., Botrè, F., Vesin, J. M., & Faiss, R. (2023). Prediction of plasma volume and total hemoglobin mass with machine learning. Physiological Reports, 11(19), e15834. https://doi.org/10.14814/phy2.15834